Factor out VisualBench timing code into a helper class

BUG=skia:

Review URL: https://codereview.chromium.org/1375363003

10 files changed, 292 insertions, 218 deletions

diff --git a/bench/BigPathBench.cpp b/bench/BigPathBench.cpp
index 0ed8fc51e3..f39fd73693 100644
--- a/bench/BigPathBench.cpp
+++ b/bench/BigPathBench.cpp
@@ -8,11 +8,7 @@
 #include "Benchmark.h"
 #include "SkCanvas.h"
 #include "SkPath.h"
-#include "SkString.h"
-
-static void make_path(SkPath& path) {
-    #include "BigPathBench.inc"
-}
+#include "sk_tool_utils.h"
 
 enum Align {
     kLeft_Align,
@@ -47,7 +43,7 @@ protected:
     }
 
     void onDelayedSetup() override {
-        make_path(fPath);
+        sk_tool_utils::make_big_path(fPath);
     }
 
     void onDraw(int loops, SkCanvas* canvas) override {
diff --git a/bench/BigPathBench.inc b/tools/BigPathBench.inc
index d2d147ff40..c52bd32353 100644
--- a/bench/BigPathBench.inc
+++ b/tools/BigPathBench.inc
@@ -1,3 +1,4 @@
+// Inspired by crbug.com/455429
 path.moveTo(-1113, 81);
 path.lineTo(-1113, 83);
 path.lineTo(-1113, 82);
diff --git a/tools/VisualBench/TimingStateMachine.cpp b/tools/VisualBench/TimingStateMachine.cpp
new file mode 100644
index 0000000000..c7f2f13470
--- /dev/null
+++ b/tools/VisualBench/TimingStateMachine.cpp
@@ -0,0 +1,122 @@
+/*
+ * Copyright 2015 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ */
+
+#include "TimingStateMachine.h"
+
+#include "SkCanvas.h"
+#include "SkCommandLineFlags.h"
+
+DEFINE_int32(gpuFrameLag, 5, "Overestimate of maximum number of frames GPU is allowed to lag.");
+DEFINE_int32(frames, 5, "Number of frames of each skp to render per sample.");
+DEFINE_double(loopMs, 5, "Each benchmark will be tuned until it takes loopsMs millseconds.");
+
+TimingStateMachine::TimingStateMachine()
+    : fCurrentFrame(0)
+    , fLoops(1)
+    , fLastMeasurement(0.)
+    , fState(kPreWarmLoopsPerCanvasPreDraw_State) {
+}
+
+TimingStateMachine::ParentEvents TimingStateMachine::nextFrame(SkCanvas* canvas,
+                                                               Benchmark* benchmark) {
+    switch (fState) {
+        case kPreWarmLoopsPerCanvasPreDraw_State:
+            return this->perCanvasPreDraw(canvas, benchmark, kPreWarmLoops_State);
+        case kPreWarmLoops_State:
+            return this->preWarm(kTuneLoops_State);
+        case kTuneLoops_State:
+            return this->tuneLoops();
+        case kPreWarmTimingPerCanvasPreDraw_State:
+            return this->perCanvasPreDraw(canvas, benchmark, kPreWarmTiming_State);
+        case kPreWarmTiming_State:
+            return this->preWarm(kTiming_State);
+        case kTiming_State:
+            return this->timing(canvas, benchmark);
+    }
+    SkFAIL("Incomplete switch\n");
+    return kTiming_ParentEvents;
+}
+
+inline void TimingStateMachine::nextState(State nextState) {
+    fState = nextState;
+}
+
+TimingStateMachine::ParentEvents TimingStateMachine::perCanvasPreDraw(SkCanvas* canvas,
+                                                                      Benchmark* benchmark,
+                                                                      State nextState) {
+    benchmark->perCanvasPreDraw(canvas);
+    benchmark->preDraw(canvas);
+    fCurrentFrame = 0;
+    this->nextState(nextState);
+    return kTiming_ParentEvents;
+}
+
+TimingStateMachine::ParentEvents TimingStateMachine::preWarm(State nextState) {
+    if (fCurrentFrame >= FLAGS_gpuFrameLag) {
+        // we currently time across all frames to make sure we capture all GPU work
+        this->nextState(nextState);
+        fCurrentFrame = 0;
+        fTimer.start();
+    } else {
+        fCurrentFrame++;
+    }
+    return kTiming_ParentEvents;
+}
+
+inline double TimingStateMachine::elapsed() {
+    fTimer.end();
+    return fTimer.fWall;
+}
+
+void TimingStateMachine::resetTimingState() {
+    fCurrentFrame = 0;
+    fTimer = WallTimer();
+}
+
+inline TimingStateMachine::ParentEvents TimingStateMachine::tuneLoops() {
+    if (1 << 30 == fLoops) {
+        // We're about to wrap.  Something's wrong with the bench.
+        SkDebugf("InnerLoops wrapped\n");
+        fLoops = 1;
+        return kTiming_ParentEvents;
+    } else {
+        double elapsedMs = this->elapsed();
+        if (elapsedMs > FLAGS_loopMs) {
+            this->nextState(kPreWarmTimingPerCanvasPreDraw_State);
+        } else {
+            fLoops *= 2;
+            this->nextState(kPreWarmLoops_State);
+        }
+        this->resetTimingState();
+        return kReset_ParentEvents;
+    }
+}
+
+void TimingStateMachine::recordMeasurement() {
+    fLastMeasurement = this->elapsed() / (FLAGS_frames * fLoops);
+}
+
+void TimingStateMachine::nextBenchmark(SkCanvas* canvas, Benchmark* benchmark) {
+    benchmark->postDraw(canvas);
+    benchmark->perCanvasPostDraw(canvas);
+    fLoops = 1;
+    this->nextState(kPreWarmLoopsPerCanvasPreDraw_State);
+}
+
+inline TimingStateMachine::ParentEvents TimingStateMachine::timing(SkCanvas* canvas,
+                                                                   Benchmark* benchmark) {
+    if (fCurrentFrame >= FLAGS_frames) {
+        this->recordMeasurement();
+        this->resetTimingState();
+        this->nextState(kPreWarmTimingPerCanvasPreDraw_State);
+        return kTimingFinished_ParentEvents;
+    } else {
+        fCurrentFrame++;
+        return kTiming_ParentEvents;
+    }
+}
+
diff --git a/tools/VisualBench/TimingStateMachine.h b/tools/VisualBench/TimingStateMachine.h
new file mode 100644
index 0000000000..69ea24337f
--- /dev/null
+++ b/tools/VisualBench/TimingStateMachine.h
@@ -0,0 +1,103 @@
+/*
+ * Copyright 2015 Google Inc.
+ *
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ *
+ */
+
+#ifndef TimingStateMachine_DEFINED
+#define TimingStateMachine_DEFINED
+
+#include "Benchmark.h"
+#include "SkTArray.h"
+#include "Timer.h"
+
+class SkCanvas;
+
+/*
+ * Manages a timer via a state machine.  Can be used by modules to time benchmarks
+ *
+ * Clients call nextFrame, and must handle any requests from the timing state machine, specifically
+ * to reset.  When kTimingFinished_ParentEvents is returned, then lastMeasurement() will return the
+ * timing and loops() will return the number of loops used to time.
+ *
+ * A client may continue timing the same benchmark indefinitely.  To advance to the next
+ * benchmark, the client should call nextBenchmark.
+ */
+class TimingStateMachine {
+public:
+    TimingStateMachine();
+
+    enum ParentEvents {
+        kReset_ParentEvents,
+        kTiming_ParentEvents,
+        kTimingFinished_ParentEvents,// This implies parent can read lastMeasurement() and must
+                                     // reset
+    };
+
+    ParentEvents nextFrame(SkCanvas* canvas, Benchmark* benchmark);
+
+    /*
+     * The caller should call this when they are ready to move to the next benchmark.  The caller
+     * must call this with the *last* benchmark so post draw hooks can be invoked
+     */
+    void nextBenchmark(SkCanvas*, Benchmark*);
+
+
+    /*
+     * When TimingStateMachine returns kTimingFinished_ParentEvents, then the owner can call
+     * lastMeasurement() to get the time
+     */
+    double lastMeasurement() const { return fLastMeasurement; }
+
+    int loops() const { return fLoops; }
+
+private:
+    /*
+     * The heart of the timing state machine is an event driven timing loop.
+     * kPreWarmLoopsPerCanvasPreDraw_State:  Before we begin timing, Benchmarks have a hook to
+     *                                       access the canvas.  Then we prewarm before the autotune
+     *                                       loops step.
+     * kPreWarmLoops_State:                  We prewarm the gpu before auto tuning to enter a steady
+     *                                       work state
+     * kTuneLoops_State:                     Then we tune the loops of the benchmark to ensure we
+     *                                       are doing a measurable amount of work
+     * kPreWarmTimingPerCanvasPreDraw_State: Because reset the context after tuning loops to ensure
+     *                                       coherent state, we need to give the benchmark
+     *                                       another hook
+     * kPreWarmTiming_State:                 We prewarm the gpu again to enter a steady state
+     * kTiming_State:                        Finally we time the benchmark.  When finished timing
+     *                                       if we have enough samples then we'll start the next
+     *                                       benchmark in the kPreWarmLoopsPerCanvasPreDraw_State.
+     *                                       otherwise, we enter the
+     *                                       kPreWarmTimingPerCanvasPreDraw_State for another sample
+     *                                       In either case we reset the context.
+     */
+    enum State {
+        kPreWarmLoopsPerCanvasPreDraw_State,
+        kPreWarmLoops_State,
+        kTuneLoops_State,
+        kPreWarmTimingPerCanvasPreDraw_State,
+        kPreWarmTiming_State,
+        kTiming_State,
+    };
+
+    inline void nextState(State);
+    ParentEvents perCanvasPreDraw(SkCanvas*, Benchmark*, State);
+    ParentEvents preWarm(State nextState);
+    inline ParentEvents tuneLoops();
+    inline ParentEvents timing(SkCanvas*, Benchmark*);
+    inline double elapsed();
+    void resetTimingState();
+    void postDraw(SkCanvas*, Benchmark*);
+    void recordMeasurement();
+
+    int fCurrentFrame;
+    int fLoops;
+    double fLastMeasurement;
+    WallTimer fTimer;
+    State fState;
+};
+
+#endif
diff --git a/tools/VisualBench/VisualBenchmarkStream.cpp b/tools/VisualBench/VisualBenchmarkStream.cpp
index 9e1ce36fbc..c520eeed05 100644
--- a/tools/VisualBench/VisualBenchmarkStream.cpp
+++ b/tools/VisualBench/VisualBenchmarkStream.cpp
@@ -10,8 +10,10 @@
 #include "CpuWrappedBenchmark.h"
 #include "GMBench.h"
 #include "SkOSFile.h"
+#include "SkPath.h"
 #include "SkPictureRecorder.h"
 #include "SkStream.h"
+#include "sk_tool_utils.h"
 #include "VisualSKPBench.h"
 
 DEFINE_bool(cpu, false, "Run in CPU mode?");
@@ -26,12 +28,41 @@ DEFINE_string2(match, m, nullptr,
                "it is skipped unless some list entry starts with ~");
 DEFINE_string(skps, "skps", "Directory to read skps from.");
 
+// We draw a big nonAA path to warmup the gpu / cpu
+#include "SkPerlinNoiseShader.h"
+class WarmupBench : public Benchmark {
+public:
+    WarmupBench() {
+        sk_tool_utils::make_big_path(fPath);
+    }
+private:
+    const char* onGetName() override { return "warmupbench"; }
+    void onDraw(int loops, SkCanvas* canvas) override {
+        // We draw a big path to warm up the cpu, and then use perlin noise shader to warm up the
+        // gpu
+        SkPaint paint;
+        paint.setStyle(SkPaint::kStroke_Style);
+        paint.setStrokeWidth(2);
+
+        SkPaint perlinPaint;
+        perlinPaint.setShader(SkPerlinNoiseShader::CreateTurbulence(0.1f, 0.1f, 1, 0,
+                                                                    nullptr))->unref();
+        SkRect rect = SkRect::MakeLTRB(0., 0., 400., 400.);
+        for (int i = 0; i < loops; i++) {
+            canvas->drawPath(fPath, paint);
+            canvas->drawRect(rect, perlinPaint);
+        }
+    }
+    SkPath fPath;
+};
+
 VisualBenchmarkStream::VisualBenchmarkStream()
     : fBenches(BenchRegistry::Head())
     , fGMs(skiagm::GMRegistry::Head())
     , fSourceType(nullptr)
     , fBenchType(nullptr)
-    , fCurrentSKP(0) {
+    , fCurrentSKP(0)
+    , fIsWarmedUp(false) {
     for (int i = 0; i < FLAGS_skps.count(); i++) {
         if (SkStrEndsWith(FLAGS_skps[i], ".skp")) {
             fSKPs.push_back() = FLAGS_skps[i];
@@ -67,7 +98,13 @@ bool VisualBenchmarkStream::ReadPicture(const char* path, SkAutoTUnref<SkPicture
 }
 
 Benchmark* VisualBenchmarkStream::next() {
+    if (!fIsWarmedUp) {
+        fIsWarmedUp = true;
+        return new WarmupBench;
+    }
+
     Benchmark* bench;
+
     // skips non matching benches
     while ((bench = this->innerNext()) &&
            (SkCommandLineFlags::ShouldSkip(FLAGS_match, bench->getUniqueName()) ||
diff --git a/tools/VisualBench/VisualBenchmarkStream.h b/tools/VisualBench/VisualBenchmarkStream.h
index e8d2cbe524..9e5d4590cb 100644
--- a/tools/VisualBench/VisualBenchmarkStream.h
+++ b/tools/VisualBench/VisualBenchmarkStream.h
@@ -34,6 +34,7 @@ private:
     const char* fSourceType;  // What we're benching: bench, GM, SKP, ...
     const char* fBenchType;   // How we bench it: micro, playback, ...
     int fCurrentSKP;
+    bool fIsWarmedUp;
 };
 
 #endif
diff --git a/tools/VisualBench/VisualLightweightBenchModule.cpp b/tools/VisualBench/VisualLightweightBenchModule.cpp
index cdb9318c4b..fa99caa1fb 100644
--- a/tools/VisualBench/VisualLightweightBenchModule.cpp
+++ b/tools/VisualBench/VisualLightweightBenchModule.cpp
@@ -25,18 +25,13 @@ __SK_FORCE_IMAGE_DECODER_LINKING;
 
 // Between samples we reset context
 // Between frames we swap buffers
-
-DEFINE_int32(maxWarmupFrames, 100, "maxmium frames to try and tune for sane timings");
-DEFINE_int32(gpuFrameLag, 5, "Overestimate of maximum number of frames GPU allows to lag.");
-DEFINE_int32(samples, 10, "Number of times to time each skp.");
-DEFINE_int32(frames, 5, "Number of frames of each skp to render per sample.");
-DEFINE_double(loopMs, 5, "Target loop time in millseconds.");
 DEFINE_bool2(verbose, v, false, "enable verbose output from the test driver.");
 DEFINE_string(outResultsFile, "", "If given, write results here as JSON.");
 DEFINE_string(key, "",
               "Space-separated key/value pairs to add to JSON identifying this builder.");
 DEFINE_string(properties, "",
               "Space-separated key/value pairs to add to JSON identifying this run.");
+DEFINE_int32(samples, 10, "Number of times to time each skp.");
 
 static SkString humanize(double ms) {
     if (FLAGS_verbose) {
@@ -47,34 +42,8 @@ static SkString humanize(double ms) {
 
 #define HUMANIZE(time) humanize(time).c_str()
 
-// We draw a big nonAA path to warmup the gpu / cpu
-class WarmupBench : public Benchmark {
-public:
-    WarmupBench() {
-        make_path(fPath);
-    }
-private:
-    static void make_path(SkPath& path) {
-        #include "BigPathBench.inc"
-    }
-    const char* onGetName() override { return "warmupbench"; }
-    void onDraw(int loops, SkCanvas* canvas) override {
-        SkPaint paint;
-        paint.setStyle(SkPaint::kStroke_Style);
-        paint.setStrokeWidth(2);
-        for (int i = 0; i < loops; i++) {
-            canvas->drawPath(fPath, paint);
-        }
-    }
-    SkPath fPath;
-};
-
 VisualLightweightBenchModule::VisualLightweightBenchModule(VisualBench* owner)
     : fCurrentSample(0)
-    , fCurrentFrame(0)
-    , fLoops(1)
-    , fState(kWarmup_State)
-    , fBenchmark(nullptr)
     , fOwner(SkRef(owner))
     , fResults(new ResultsWriter) {
     fBenchmarkStream.reset(new VisualBenchmarkStream);
@@ -106,7 +75,7 @@ VisualLightweightBenchModule::VisualLightweightBenchModule(VisualBench* owner)
 }
 
 inline void VisualLightweightBenchModule::renderFrame(SkCanvas* canvas) {
-    fBenchmark->draw(fLoops, canvas);
+    fBenchmark->draw(fTSM.loops(), canvas);
     canvas->flush();
     fOwner->present();
 }
@@ -140,7 +109,7 @@ void VisualLightweightBenchModule::printStats() {
         SkDebugf("%4d/%-4dMB\t%d\t%s\t%s\t%s\t%s\t%.0f%%\t%s\t%s\n",
                  sk_tools::getCurrResidentSetSizeMB(),
                  sk_tools::getMaxResidentSetSizeMB(),
-                 fLoops,
+                 fTSM.loops(),
                  HUMANIZE(stats.min),
                  HUMANIZE(stats.median),
                  HUMANIZE(stats.mean),
@@ -152,12 +121,6 @@ void VisualLightweightBenchModule::printStats() {
 }
 
 bool VisualLightweightBenchModule::advanceRecordIfNecessary(SkCanvas* canvas) {
-    if (!fBenchmark && fState == kWarmup_State) {
-        fOwner->clear(canvas, SK_ColorWHITE, 2);
-        fBenchmark.reset(new WarmupBench);
-        return true;
-    }
-
     if (fBenchmark) {
         return true;
     }
@@ -178,49 +141,6 @@ bool VisualLightweightBenchModule::advanceRecordIfNecessary(SkCanvas* canvas) {
     return true;
 }
 
-inline void VisualLightweightBenchModule::nextState(State nextState) {
-    fState = nextState;
-}
-
-void VisualLightweightBenchModule::perCanvasPreDraw(SkCanvas* canvas, State nextState) {
-    fBenchmark->perCanvasPreDraw(canvas);
-    fBenchmark->preDraw(canvas);
-    fCurrentFrame = 0;
-    this->nextState(nextState);
-}
-
-void VisualLightweightBenchModule::warmup(SkCanvas* canvas) {
-    if (fCurrentFrame >= FLAGS_maxWarmupFrames) {
-        this->nextState(kPreWarmLoopsPerCanvasPreDraw_State);
-        fBenchmark.reset(nullptr);
-        this->resetTimingState();
-        fLoops = 1;
-    } else {
-        bool isEven = (fCurrentFrame++ % 2) == 0;
-        if (isEven) {
-            fTimer.start();
-        } else {
-            double elapsedMs = this->elapsed();
-            if (elapsedMs < FLAGS_loopMs) {
-                fLoops *= 2;
-            }
-            fTimer = WallTimer();
-            fOwner->reset();
-        }
-    }
-}
-
-void VisualLightweightBenchModule::preWarm(State nextState) {
-    if (fCurrentFrame >= FLAGS_gpuFrameLag) {
-        // we currently time across all frames to make sure we capture all GPU work
-        this->nextState(nextState);
-        fCurrentFrame = 0;
-        fTimer.start();
-    } else {
-        fCurrentFrame++;
-    }
-}
-
 void VisualLightweightBenchModule::draw(SkCanvas* canvas) {
     if (!this->advanceRecordIfNecessary(canvas)) {
         SkDebugf("Exiting VisualBench successfully\n");
@@ -228,93 +148,25 @@ void VisualLightweightBenchModule::draw(SkCanvas* canvas) {
         return;
     }
     this->renderFrame(canvas);
-    switch (fState) {
-        case kWarmup_State: {
-            this->warmup(canvas);
-            break;
-        }
-        case kPreWarmLoopsPerCanvasPreDraw_State: {
-            this->perCanvasPreDraw(canvas, kPreWarmLoops_State);
-            break;
-        }
-        case kPreWarmLoops_State: {
-            this->preWarm(kTuneLoops_State);
-            break;
-        }
-        case kTuneLoops_State: {
-            this->tuneLoops();
-            break;
-        }
-        case kPreWarmTimingPerCanvasPreDraw_State: {
-            this->perCanvasPreDraw(canvas, kPreWarmTiming_State);
+    TimingStateMachine::ParentEvents event = fTSM.nextFrame(canvas, fBenchmark);
+    switch (event) {
+        case TimingStateMachine::kReset_ParentEvents:
+            fOwner->reset();
             break;
-        }
-        case kPreWarmTiming_State: {
-            this->preWarm(kTiming_State);
+        case TimingStateMachine::kTiming_ParentEvents:
             break;
-        }
-        case kTiming_State: {
-            this->timing(canvas);
+        case TimingStateMachine::kTimingFinished_ParentEvents:
+            fOwner->reset();
+            fRecords.back().fMeasurements.push_back(fTSM.lastMeasurement());
+            if (++fCurrentSample > FLAGS_samples) {
+                this->printStats();
+                fTSM.nextBenchmark(canvas, fBenchmark);
+                fCurrentSample = 0;
+                fBenchmark.reset(nullptr);
+            }
             break;
-        }
-    }
-}
-
-inline double VisualLightweightBenchModule::elapsed() {
-    fTimer.end();
-    return fTimer.fWall;
-}
-
-void VisualLightweightBenchModule::resetTimingState() {
-    fCurrentFrame = 0;
-    fTimer = WallTimer();
-    fOwner->reset();
-}
-
-inline void VisualLightweightBenchModule::tuneLoops() {
-    if (1 << 30 == fLoops) {
-        // We're about to wrap.  Something's wrong with the bench.
-        SkDebugf("InnerLoops wrapped\n");
-        fLoops = 1;
-    } else {
-        double elapsedMs = this->elapsed();
-        if (elapsedMs > FLAGS_loopMs) {
-            this->nextState(kPreWarmTimingPerCanvasPreDraw_State);
-        } else {
-            fLoops *= 2;
-            this->nextState(kPreWarmLoops_State);
-        }
-        this->resetTimingState();
     }
-}
 
-void VisualLightweightBenchModule::recordMeasurement() {
-    double measurement = this->elapsed() / (FLAGS_frames * fLoops);
-    fRecords.back().fMeasurements.push_back(measurement);
-}
-
-void VisualLightweightBenchModule::postDraw(SkCanvas* canvas) {
-    fBenchmark->postDraw(canvas);
-    fBenchmark->perCanvasPostDraw(canvas);
-    fBenchmark.reset(nullptr);
-    fCurrentSample = 0;
-    fLoops = 1;
-}
-
-inline void VisualLightweightBenchModule::timing(SkCanvas* canvas) {
-    if (fCurrentFrame >= FLAGS_frames) {
-        this->recordMeasurement();
-        if (fCurrentSample++ >= FLAGS_samples) {
-            this->printStats();
-            this->postDraw(canvas);
-            this->nextState(kPreWarmLoopsPerCanvasPreDraw_State);
-        } else {
-            this->nextState(kPreWarmTimingPerCanvasPreDraw_State);
-        }
-        this->resetTimingState();
-    } else {
-        fCurrentFrame++;
-    }
 }
 
 bool VisualLightweightBenchModule::onHandleChar(SkUnichar c) {
diff --git a/tools/VisualBench/VisualLightweightBenchModule.h b/tools/VisualBench/VisualLightweightBenchModule.h
index 1a30875c61..ffa109dc49 100644
--- a/tools/VisualBench/VisualLightweightBenchModule.h
+++ b/tools/VisualBench/VisualLightweightBenchModule.h
@@ -13,7 +13,7 @@
 
 #include "ResultsWriter.h"
 #include "SkPicture.h"
-#include "Timer.h"
+#include "TimingStateMachine.h"
 #include "VisualBench.h"
 #include "VisualBenchmarkStream.h"
 
@@ -32,65 +32,21 @@ public:
     bool onHandleChar(SkUnichar c) override;
 
 private:
-    /*
-     * The heart of visual bench is an event driven timing loop.
-     * kWarmup_State:                        We run a dummy bench to let things settle on startup
-     * kPreWarmLoopsPerCanvasPreDraw_State:  Before we begin timing, Benchmarks have a hook to
-     *                                       access the canvas.  Then we prewarm before the autotune
-     *                                       loops step.
-     * kPreWarmLoops_State:                  We prewarm the gpu before auto tuning to enter a steady
-     *                                       work state
-     * kTuneLoops_State:                     Then we tune the loops of the benchmark to ensure we
-     *                                       are doing a measurable amount of work
-     * kPreWarmTimingPerCanvasPreDraw_State: Because reset the context after tuning loops to ensure
-     *                                       coherent state, we need to give the benchmark
-     *                                       another hook
-     * kPreWarmTiming_State:                 We prewarm the gpu again to enter a steady state
-     * kTiming_State:                        Finally we time the benchmark.  When finished timing
-     *                                       if we have enough samples then we'll start the next
-     *                                       benchmark in the kPreWarmLoopsPerCanvasPreDraw_State.
-     *                                       otherwise, we enter the
-     *                                       kPreWarmTimingPerCanvasPreDraw_State for another sample
-     *                                       In either case we reset the context.
-     */
-    enum State {
-        kWarmup_State,
-        kPreWarmLoopsPerCanvasPreDraw_State,
-        kPreWarmLoops_State,
-        kTuneLoops_State,
-        kPreWarmTimingPerCanvasPreDraw_State,
-        kPreWarmTiming_State,
-        kTiming_State,
-    };
     void setTitle();
     bool setupBackend();
     void setupRenderTarget();
     void printStats();
     bool advanceRecordIfNecessary(SkCanvas*);
     inline void renderFrame(SkCanvas*);
-    inline void nextState(State);
-    void perCanvasPreDraw(SkCanvas*, State);
-    void preWarm(State nextState);
-    inline void tuneLoops();
-    inline void timing(SkCanvas*);
-    inline double elapsed();
-    void resetTimingState();
-    void postDraw(SkCanvas*);
-    void recordMeasurement();
-    void warmup(SkCanvas* canvas);
 
     struct Record {
         SkTArray<double> fMeasurements;
     };
-
     int fCurrentSample;
-    int fCurrentFrame;
-    int fLoops;
     SkTArray<Record> fRecords;
-    WallTimer fTimer;
-    State fState;
     SkAutoTDelete<VisualBenchmarkStream> fBenchmarkStream;
     SkAutoTUnref<Benchmark> fBenchmark;
+    TimingStateMachine fTSM;
 
     // support framework
     SkAutoTUnref<VisualBench> fOwner;
diff --git a/tools/sk_tool_utils.cpp b/tools/sk_tool_utils.cpp
index aab068cbd7..710c204271 100644
--- a/tools/sk_tool_utils.cpp
+++ b/tools/sk_tool_utils.cpp
@@ -345,4 +345,8 @@ void create_tetra_normal_map(SkBitmap* bm, const SkIRect& dst) {
     }
 }
 
+void make_big_path(SkPath& path) {
+    #include "BigPathBench.inc"
+}
+
 }  // namespace sk_tool_utils
diff --git a/tools/sk_tool_utils.h b/tools/sk_tool_utils.h
index b20c60ba76..e0ad1df3ec 100644
--- a/tools/sk_tool_utils.h
+++ b/tools/sk_tool_utils.h
@@ -17,6 +17,7 @@
 class SkBitmap;
 class SkCanvas;
 class SkPaint;
+class SkPath;
 class SkShader;
 class SkTestFont;
 class SkTextBlobBuilder;
@@ -133,6 +134,7 @@ namespace sk_tool_utils {
 
     void create_tetra_normal_map(SkBitmap* bm, const SkIRect& dst);
 
+    void make_big_path(SkPath& path);
 }  // namespace sk_tool_utils
 
 #endif  // sk_tool_utils_DEFINED